5.electronics 4.antennas 6.communications 3.site layout 8...
TRANSCRIPT
Exploring Wavelet Techniques���for Filtering and Triggering
John Kelley RU Nijmegen
26 October 2009
Auger Engineering Radio ArrayAERA
Jörg R. Hörandel, Radboud University Nijmegen for
Auger Engineering Radio ArrayReview, Malargüe, April 2009
1.Introduction
2.Physics case
3.Site layout
4.Antennas
5.Electronics
6.Communications
7.Power harvesting and EMC
8.DAQ
9.Risk analysis
10.Management
Previous Work!
• This work is basically at the same stage… hopefully motivate us to continue
26.10.2009 J. Kelley, Auger NL meeting 2
Ansatz to a wavelet analysis of radio signals
A.M. van den Berg1, J. Coppens
2,3, S. Harmsma
1,3, C. Timmermans
3
November 26, 2007
1KVI, University of Groningen, the Netherlands.
2IMAPP, Radboud University Nijmegen, the Netherlands.
3NIKHEF Nijmegen and Amsterdam, the Netherlands.
1 Introduction
For the identification of radio signals induced by extensive air showers filtering tech-
niques will be needed to increase the signal-to-noise ratio. A denoising technique based
on wavelet analysis is described and applied for a radio event. The radio event used
in this analysis is associated with CDAS event number 3493116 and obtained with the
test setup described in Ref. [1]. The radio event was detected using an Logarithmic
Periodic Dipole Antenna (LPDA) with its associated low-noise amplifier [2], which has
a gain of 21 dB. The antennas are connected to the radio readout system using 160 m
long RG213 cables which have an attenuation at 60 MHz of about 8 dB [3]. To reduce
signals from strong (radio and TV) transmitters, two passive filters were used: a pass
filter (Mini-circuits SBP-60), which has its pass band between 50 and 70 MHz, and
a 70 MHz low-pass filter (Mini-circuits SLP-70). After this filtering stage the signals
were fed into a 31.5 dB flat-band amplifier (Mini-circuits ZKL-2) and digitized using
a 400 MS s−1
12 bits receiver developed by NIKHEF. The present event was recorded
using antenna 1. The LPDA has two planes, one plane directed along the EW di-
rection, the other one along the NS direction. After this initial analysis a ‘random’
sample of events was used to study different digital trigger conditions.
Because of the narrow pass band, a signal in the time domain resembles very much
a sine wave multiplied with a bell-shaped function. The frequency of the sine wave
corresponds to the maximum frequency in the pass filter and the position of the bell-
shaped function in the time domain reflects the (narrow) time position of the induced
pulse. Typically, the length of a pulse is around 10 - 100 ns. As discussed amongst
others by Torrence and Compo [4], a wavelet analysis provides an efficient analysis
method for such time-and-frequency localized pulses, as compared to e.g. a windowed
Fourier transform.
The present document describes a wavelet analysis to the measured pulse, showing
that indeed wavelet analysis can be used to reduce the noise level significantly. The
analysis is based on the software [5] made available by Torrence and Compo [4].
2 Analysis
The wavelet basis used in the present analysis is of the Morlet type. For a wavelet a
scale s can be defined. In the time domain this scale can be regarded as the duration of
the wavelet, i.e. a slow oscillation for a large-scale fluctuation, versus rapid oscillations
for small scales. Mathematically, this function is given as a plane wave multiplied with
1
Why Wavelets?
• A decomposition of a time series that provides both spectral and temporal information – Possibly better than DFT for transient analysis
• Successful use in other fields for weak transient extraction – Radar reflections (see e.g. Ehara, Sasasi, and Mori 1994) – GRB light curves (see e.g. thesis by N. Butler, MIT 2003)
• Transforms can be fast – Possible implementation in FPGA for trigger
26.10.2009 3 J. Kelley, Auger NL meeting
Discrete Wavelet Transform
• Basis functions – finite in time – scaled and shifted
versions of a “mother wavelet”
• Decomposition results – wavelet coefficients – time on one axis – period scale (often
factors of 2) on the other
26.10.2009 4
Haar
Daubechies 4
Morlet
J. Kelley, Auger NL meeting
Visualization of Haar DWT
26.10.2009 5 J. Kelley, Auger NL meeting
= c00 ×
c10 c11
c20 c21 c22 c23
+
+
+
+ …
c30…c37
Visualization of Haar DWT
26.10.2009 6 J. Kelley, Auger NL meeting
= c00 ×
c10 c11
c20 c21 c22 c23
+
+
+
+ …
c30…c37
Example Radio Pulse (2007)
26.10.2009 J. Kelley, Auger NL meeting 7
s)µTime (2 3 4 5 6 7
V)
µV
olt
ag
e (
-40
-30
-20
-10
0
10
20
30
40
50
s)µTime (2 3 4 5 6 7
) s
P/T
2P
eri
od
scale
(lo
g
2
4
6
8
10
12
-80
-60
-40
-20
0
20
40
60
Event 3388350, pole 2, NS Daubechies4 DWT coefficients
Projection (200 MHz)
26.10.2009 J. Kelley, Auger NL meeting 8
Projection (100 MHz)
26.10.2009 J. Kelley, Auger NL meeting 9
Projection (50 MHz)
26.10.2009 J. Kelley, Auger NL meeting 10
Projection (25 MHz)
26.10.2009 J. Kelley, Auger NL meeting 11
Projection (12.5 MHz)
26.10.2009 J. Kelley, Auger NL meeting 12
Projection (6.3 MHz)
26.10.2009 J. Kelley, Auger NL meeting 13
Projection (3.1 MHz)
26.10.2009 J. Kelley, Auger NL meeting 14
Another Example
26.10.2009 J. Kelley, Auger NL meeting 15
s)µTime (2 3 4 5 6 7
V)
µV
olt
ag
e (
-20
-10
0
10
20
Event 3382961, pole 1, NS
TTL noise cosmic
s)µTime (2 3 4 5 6 7
) s
P/T
2P
eri
od
sc
ale
(lo
g2
4
6
8
10
12
-40
-30
-20
-10
0
10
20
30
40
50
coefficients (cij)
Another Example
26.10.2009 J. Kelley, Auger NL meeting 16
s)µTime (2 3 4 5 6 7
V)
µV
olt
ag
e (
-20
-10
0
10
20
Event 3382961, pole 1, NS
TTL noise cosmic
power (cij2)
Filtering
• Hard threshold on wavelet coefficients (or power) can keep pulses but reject white noise
• Coefficient threshold by Donoho & Johnstone:
• All coefficients below τ in magnitude set to 0
• Inverse transform to reconstruct filtered time series 26.10.2009 J. Kelley, Auger NL meeting 17
τ = δ�
lnNsamp (11)
δ =median(|cHF,j |)
0.6745(12)
2
Filtering example
26.10.2009 J. Kelley, Auger NL meeting 18
s)µTime (2 3 4 5 6 7
V)
µV
olt
ag
e (
-20
-10
0
10
20
Event 3382961, pole 1, NS
s)µTime (2 3 4 5 6 7
V)
µV
olt
ag
e (
-20
-10
0
10
20
cij threshold = 2 τDonoho
Filtering example
26.10.2009 J. Kelley, Auger NL meeting 19
s)µTime (2 3 4 5 6 7
V)
µV
olt
ag
e (
-20
-10
0
10
20
Event 3382961, pole 1, NS cij threshold = 4 τDonoho
s)µTime (2 3 4 5 6 7
V)
µV
olt
ag
e (
-30
-20
-10
0
10
20
30
Filtering example
26.10.2009 J. Kelley, Auger NL meeting 20
s)µTime (2 3 4 5 6 7
V)
µV
olt
ag
e (
-20
-10
0
10
20
Event 3382961, pole 1, NS cij threshold = 6 τDonoho
s)µTime (2 3 4 5 6 7
V)
µV
olt
ag
e (
-20
-10
0
10
20
Filtering example
26.10.2009 J. Kelley, Auger NL meeting 21
s)µTime (2 3 4 5 6 7
V)
µV
olt
ag
e (
-20
-10
0
10
20
Event 3382961, pole 1, NS cij threshold = 8 τDonoho
s)µTime (2 3 4 5 6 7
V)
µV
olt
ag
e (
-20
-10
0
10
20
Difficulties
• Filtering alone doesn’t distinguish between good and bad pulses
• Transform + filtering + inverse probably too much for L1 trigger in FPGA – Possible exception: short timescale Haar
• Could be useful for L2 or offline pulse extraction
26.10.2009 J. Kelley, Auger NL meeting 22
Trigger?
26.10.2009 J. Kelley, Auger NL meeting 23
rolling integral over 25-100 MHz coefficient bins (abs. value)
s)µTime (0 1 2 3 4 5 6 7 8 9 10
V)
µV
olt
ag
e (
-40
-30
-20
-10
0
10
20
30
40
50
s)µTime (2 3 4 5 6 7
) s
P/T
2P
eri
od
scale
(lo
g
2
4
6
8
10
12
-60
-40
-20
0
20
40
60
Rolling Sum
• Looks promising — but triggering on golden events trivial
• Will likely still trigger on TTL pulse
• May be good for power line noise — still looking
• Don’t know about pulse trains yet
26.10.2009 J. Kelley, Auger NL meeting 24
s)µTime (0 1 2 3 4 5 6 7 8 9
| (a
.u.)
ij |
c!
1000
2000
3000
4000
5000
6000
Summary
• In principle, wavelet analysis is appropriate – provides temporal and spectral information
• Threshold filtering can pull pulses out of noise – could also use for data compression, but impact on spectral
information?
• Haar wavelet transform feasible in FPGA
• A wavelet L1 trigger is possible but needs much more work on discrimination power – ratio in different frequency bands?
26.10.2009 J. Kelley, Auger NL meeting 25